Over-refusal Evaluation on XSTest (test)
0.035Over-refusal RateClaude Sonnet 4.5
Evaluation Results
| Method | Links | |
|---|---|---|
| Claude Sonnet 4.5Safety Steering=Disabled2025.02 | 0.035 | |
| GPT-5Safety Steering=Disabled2025.02 | 0.052 | |
| Claude Sonnet 4.5Safety Steering=Enabled2025.02 | 0.074 | |
| GPT-5Safety Steering=Enabled2025.02 | 0.122 |